Professional Documents
Culture Documents
Evaluating A New Generation of Expansive Claims About Vote Manipulation
Evaluating A New Generation of Expansive Claims About Vote Manipulation
∗
about Vote Manipulation
a b c
Justin Grimmer , Michael C. Herron , and Matthew Tyler
a
Democracy and Polarization Lab, Department of Political Science, and Hoover Institution,
Stanford University. 616 Jane Stanford Way, Stanford CA 94305 ([email protected]).
b
Program in Quantitative Social Science, Dartmouth College. 6108 Silsby Hall, Hanover NH
03755 ([email protected]).
c
Postdoctoral Fellow at the Democracy and Polarization Lab and the American National Election
Studies, Stanford University. 616 Jane Stanford Way, Stanford CA 94305 ([email protected]).
Abstract
In the wake of Donald Trump’s attempt to overturn the 2020 presidential election, a
cottage industry of conspiracy theorists has advanced ever more expansive claims of
vote manipulation, going so far as to allege that all American elections are subject to
manipulation—even in largely Republican states. In the extreme, these conspiracy
theorists argue that candidates in U.S. elections are selected rather than elected. We
evaluate two recent sets of claims about vote manipulation that allege algorithms are
used to shift votes towards preferred candidates. Even though these claims are
distinct, they fail for similar reasons. For example, both sets of claims assert that
“unnaturally” accurate predictions of election results are evidence of vote
manipulation, an allegation that is a result of predicting a variable with itself.
Furthermore, both claims make easily refuted errors in logic and data analysis and in
addition misrepresent historical election patterns. While recent claims about vote
manipulation are prima facie outlandish, their effects on policy and the public are
real. Refuting false claims about vote manipulation is essential to ensuring the
continued functioning of U.S. elections and American democracy more generally.
∗
Grimmer and Herron served as expert witnesses in Gilbert v. Sisolak (Case No.: 22 OC 000851B), an
election contest discussed at length in this paper. Grimmer also served as an expert witness in Washington
County v. Tim Sippel (Case No.: 22 CV 07782), where the defendants unsuccessfully attempted to include
Douglas Frank as an expert on elections.
1
1 Introduction
Donald Trump’s refusal to concede the 2020 presidential election was a shock to what had
previously been bipartisan faith in the U.S. electoral system. And almost two years on,
Trump continues to claim that he won the 2020 presidential race (Hansen 2022). In
response to allegations that the 2020 election was stolen (Hasen 2022), scholars, journalists,
and the courts have resoundingly refuted a multitude of claims about illicit vote
manipulation (Cassidy 2021; Eggers et al. 2021; Danforth et al. 2022). While other
presidential elections have been the subject of allegations of malfeasance (Cottrell et al.
2018), the scale of claims made in the aftermath of 2020 election dwarfs what the United
States has previously experienced.
Despite the failure of Trump’s objections to the 2020 election, the consequences of his
unprecedented allegations continue to unfold. Notably, his claims have helped lay the
foundation for even more expansive claims of vote manipulation. Unlike Trump’s initial
post-2020 election allegations, these new claims of malfeasance indict nearly every election
in the United States, including Republican primary elections, elections in deep red states
that resulted in Republican legislative super-majorities, and even city council elections in
1
relatively small towns. In the extreme, these new conspiracies allege that winners of U.S.
elections are selected rather than elected.
In this paper, we document two high-profile examples of expansive claims of vote
manipulation and demonstrate how supposed evidence in favor of the claims fails to
establish any evidence of malfeasance. Our first example considers allegations made by the
second-place candidate in the 2022 Nevada Republican gubernatorial primary, Joey
Gilbert, who claimed that vote manipulation led to his June 2022 loss to Joseph Lombardo,
the Trump-endorsed sheriff of Clark County, Nevada. Gilbert’s claims about vote
manipulation were the foundation of an election contest filed on July 15, 2022 in Nevada’s
1
See Frank (2022b), a presentation by Douglas Frank in Idaho, and both Thayer (2022) and Corasaniti
and Berzon (2022) on allegations of vote manipulation in Grand Junction, Colorado.
2
first judicial district (Gilbert v. Sisolak, Case No.: 22 OC 000851B). Gilbert’s contest
alleged that an unspecified group of conspirators deployed an algorithm to shift votes in his
gubernatorial primary from Gilbert to Lombardo and to the incumbent Democratic
Nevada governor Steve Sisolak. The contest was dismissed on August 11, 2022, with
Carson City District Court Judge James Wilson subsequently ordering sanctions against
Gilbert and requiring him to pay attorney fees for Lombardo (Seeman 2022).
Our second example considers an expansive election conspiracy theory that is regularly
pitched to local election officials and civic groups as part of a campaign to influence the
administration of future elections (Mollenkamp et al. 2022). In particular, we examine
Douglas Frank’s claims that an algorithm determines the results of all U.S. elections. Frank’s
claims are met with a receptive audience, creating chaos for local election officials as they
prepare for the upcoming 2022 midterm election (Brown and Swaine 2022; Halpern 2022;
Hubler 2022), and they are particularly important to refute because they have been amplified
by Mike Lindell, a pillow magnate and Donald Trump supporter. Lindell uses his wealth
to push claims about vote manipulation on his eponymous television network, with movie-
length arguments that elections are stolen (Thayer 2022) and annual symposia like the
“Cyber Symposium” (Marks 2021) and the “Moment of Truth Summit.” Lindell’s television
network and his symposia have provided a platform for a cadre of election conspiracy theorists
to bolster their claims and reach a wide audience of election skeptics. Frank’s claims have
also surfaced in court proceedings, including an Oregon state-level case involving access to
voting machine information (Washington County v. Tim Sippel (Case No.: 22 CV 07782))
and a federal lawsuit alleging voter fraud throughout the state of Oregon (Thielman et al.
v. Fagan et al. (Case No.: 22 CV 01516)).
We demonstrate that Gilbert’s and Frank’s claims fundamentally fail to provide
evidence of any vote manipulation, let alone malfeasance that infects all U.S. elections.
While Gilbert’s and Frank’s theories are distinct, they fail for similar reasons. First, both
theories allege that election results were manufactured, evidence for this being the fact that
3
observed results are easy to predict with a seemingly small set of predictors (e.g., the age
distribution of voters, as we discuss in detail later in this paper). However, on closer
examination, what Gilbert and Frank consider seemingly “unnatural” predictability is
merely the result of correlating a variable with itself after applying opaque and sometimes
convoluted transformations of election data. Second, both Gilbert’s and Frank’s claims fail
due to errors in argument logic and flaws in statistical analyses. In addition, in their
theories of vote manipulation both Gilbert and Frank offer claims about historical election
data that are demonstrably false. Overall, neither the Gilbert contest nor Douglas Frank’s
expansive claims about vote manipulation provide any evidence that U.S. elections are
regularly victimized by malfeasance.
Claims like Gilbert’s and Frank’s are far from mere annoyances. Rather, expansive
claims about vote manipulation exert measurable costs on the infrastructure of American
elections. Candidates who have sufficient financial support, exemplified by Joey Gilbert,
can file lawsuits with all their attendant features, i.e., hearings, reports, and depositions, all
of which are costly in a financial sense and can, in principle, delay certification. Moreover,
proponents of election conspiracies increasingly occupy the time of elections officials. They
flood local electoral offices with public records requests (Gardner and Marley 2022) and
lead voters to ask officials about “rigged” voting machines and whether mail-in ballots are
legal (Kang 2022), complicating the jobs of local election officials. Moreover, claims of vote
manipulation often find a sympathetic audience, where allegations of election manipulation
and voter fraud can be reiterated and amplified throughout local communities.
Research in the area of election integrity is essential to provide the public, judges, and
election officials with tools to refute unsupported claims. Fair elections, a cornerstone of
democratic politics, are at risk when unsupported claims about malfeasance cause logjams
in the court system and cause voters to lose trust in democracy (Berlinski et al. 2021).
The best response to claims of vote manipulation is a careful examination of the underlying
evidence.
4
2 Vote manipulation in contemporary American
elections
5
3 “Irrefutable Geometric Proof” of vote manipulation
and claims of algorithms in Nevada’s 2022
Republican gubernatorial primary
6
4
driving (Appendix B); G. Donald Allen, a former mathematics professor at Texas A&M
University (Appendix D); and, Walter C. Daugherity, a former computer science lecturer at
5
Texas A&M (Appendix F).
The Gilbert contest alleges that candidates in an election should, within precincts, receive
roughly equivalent vote shares across methods of voting. In Nevada this would imply that
Gilbert should have received equal shares from in-person voting on Election Day, in-person
early voting, and voting via mail ballot. Examining the results from the 2022 Republican
gubernatorial primary, the Gilbert contest argued that precinct-level deviations from vote-
share equality in each mode of voting provided evidence that vote manipulation tarred this
election. On this point, the contest states that, “There is absolutely no correlation between
Gilbert’s Election Day, Early, and Mail-in Percentages across precincts,” labeling this an
“irregularity” (pp. 8-9, ¶ 29).
The Gilbert contest claims as well that Nevada’s 2022 Republican gubernatorial primary
was predictable, indicating the presence of an “illegal formula” (pp. 9-10, ¶ 32) which lead to
“geometric interference” (p. 12, ¶ 36). Solomon’s Appendix A argues that such predictability
was the product of a “Neural Network” (p. 8).
The claims made in the Gilbert contest fail for two reasons, among others. First, they
mischaracterize historical election results: in fact, there is no evidence that candidates in
Nevada elections receive equal vote shares across different methods of voting. This is
unsurprising given that different types of voters may have different preferences for, or
varied access to, alternative methods of voting. Second, what the contest argues is
allegedly surprising predictability is merely the result of using a variable to predict itself —
4
On Dr. Hemmers’s previous expert work, see lines 15-24 on p. 14 of the transcript of his July 28, 2022
deposition in Gilbert v. Sisolak.
5
While the contest’s Appendix A was not signed by Edward Solomon, the body of the contest states that
Solomon was the appendix’s author (pp. 7-8, ¶ 25) Moreover, Allen in his report identifies Appendix A as
Solomon’s work, writing in the contest’s Appendix D that, “I have reviewed, mathematically, the reports
by Edward Solomon furnished to me which mathematically analyzes (sic) the June 14, 2022, Republican
gubernatorial primacy in Clark County, Nevada, as well as other races” (p. 1, ¶ 4). Daugherity wrote
similarly of Appendix A, stating that, “I have reviewed the reports by Edward Solomon furnished to me
which mathematically analyze the June 14, 2022, Republican primary in Clark County, Nevada, as well as
other races” (Appendix F, p. 2, ¶ 6).
7
naturally, this cannot be evidence of election malfeasance.
We now consider the Gilbert contest’s claims about candidate vote shares across voting
methods and show that they do not hold.
As noted above, the Gilbert contest argues that, in fair elections, candidates should be
expected to receive (almost) identical vote shares across voting methods. The body of the
contest states that,
In a fair election, we expect a candidate that [sic] received 10% of the Election
Day Vote to get roughly 10% of the Mail-in Vote; likewise if they get 90% of
the Election Day Vote, we expect them to get 90% of the Mail-In Vote. Even if
Democrats prefer to vote by mail, that should reflect in both percentages across
the precincts, not just one of them. In other words, if we plot the election day
and mail-in percentages against each other across the precincts, they should array
themselves across a 45 degree angle y = x (Appendix A, p. 9).
Solomon grounds this claim with the following: “[E]ach candidate’s proportion of election
day, to early, to mail-in ballots, should be roughly the same, as all other candidates, in all
races. Again, this is confirmed by historical records of elections prior to 2020 and countless
simulations” (Appendix A, p. 7).
8
In his contribution to the Gilbert contest, Hemmers also reiterates the argument that
candidate vote shares should not vary by voting technology, asserting that,
In a fair election, the sum of the Early Day and Election Day votes should produce
very similar results to the Mail-in votes when the regressions analysis has a high
2
confidence (usually called R ), meaning the x-values and the y-values should be
similar (when x is 10% then y should be close to 10% as well) and not off by 25%
(Appendix B , p. 2, ¶ 5).
There is no academic literature we are aware of arguing that “fair elections” are those
that feature identical — or even similar — candidate vote shares across different voting
methods. And, to the best of our knowledge, there are no legal opinions on election fairness
that assert this. The Gilbert contest simply asserts that “fair elections” feature similar
candidate vote shares across methods of voting without any supporting scientific theory or
historical evidence.
3.1.2 Voter preferences for voting methods and state election laws imply that
candidate vote shares differ by voting method
Across the United States, voter types sort themselves into different methods of voting and
sometimes even into times of voting (Herron and Smith 2012). In the 2020 general election,
for example, relatively educated voters cast early and vote-by-mail ballots at greater rates
than voters with less education (Scherer 2021). And, in recent presidential elections, the more
educated a voter, the more likely the voter was to have supported the Democratic candidate
for president (e.g., Sides et al. 2019). If in an election voters with high (or in principle, low)
educational attainment systematically vote early and with vote-by-mail ballots, and if this
same type of voter is disproportionately Democratic or Republican, then the result will be
a correlation between voting method and candidate vote share.
The case of Texas provides an example of this phenomenon albeit based on age as
opposed to educational attainment. In Texas, voters age 65 and older are permitted to vote
absentee without an excuse, but younger voters do not have this privilege. Not surprisingly,
9
older Texas voters cast absentee ballots at greater rates than younger Texans, as noted by
Yoder et al. (2021), who study both Texas as well as Indiana, another state which provides
special absentee voting dispensation to voters 65 and up. Insofar as age is correlated with
political preferences (e.g., Ghitza et al. 2022), Texas law practically guarantees that there
is a correlation in the state between voting method and candidate vote share. Louisiana,
Mississippi, and South Carolina are three other states in which voters 65 years or older are
6
able to cast absentee ballots without excuses. Thus, the Gilbert contest’s claim that fair
elections have equal candidate vote shares across voting methods is inconsistent with the
interaction between election laws in multiple states in the country and the fact that
educational attainment and age are correlated with political preferences.
3.1.3 Nevada elections do not feature equal candidate vote shares across voting
methods
While empirical literature on American elections provides no foundation for the Gilbert
contest’s claim that fair elections feature equal candidate vote shares across methods of
voting, it could nonetheless be the case that the contest’s claim in this vein was made in a
narrow way, simply about Nevada. In light of this possibility, we show that the claim is not
supported by historical Nevada election data.
To do this, we examine precinct level election results in historical Nevada elections as well
as results from Nevada elections contemporaneous with the 2022 Republican gubernatorial
primary. For precincts in Clark County, Nevada, Figure 1 plots winning candidate vote
shares (in percentages) by method of voting for the Republican gubernatorial primaries of
2014, 2018, and 2022. Panel (a) of the figure plots Election Day and early voting vote shares
and Panel (b), Election Day and mail-in ballot vote shares. Each point in the figure’s six
6
On Louisiana, see “VOTE ABSENTEE,” Louisiana Secretary of State, available at https://www.
sos.la.gov/ElectionsAndVoting/Vote/VoteByMail/Pages/default.aspx (last accessed September 19, 2022);
on Mississippi, see “ABSENTEE VOTING FREQUENTLY ASKED QUESTIONS,” Secretary of State,
available https://www.sos.ms.gov/absentee-voting-information (last accessed September 19, 2022); and
on South Carolina, see “Absentee Voting,” South Carolina Elections Commission, available at https:
//scvotes.gov/voters/absentee-voting/ (last accessed September 19, 2022).
10
plots denotes a Clark County precinct, and points are sized proportional to total votes cast
by both relevant voting methods. Nevada election law changed in 2021 with the passage of
Assembly Bill 321, which mandates that all registered voters in Nevada receive mail ballots
7
prior to an election. That said, Figure 1 and subsequent figures in this paper treat absentee
voting and voting by mail as equivalent.
There are 17 counties in Nevada (treating Carson City as a county), and Figure 1 includes
data from Clark County only. This county, unlike the other 16 such jurisdictions in Nevada,
regularly publishes machine-readable reports that disaggregate election returns by method
of voting. Clark County is by far the most populous county in Nevada; there were 1,821,058
registered voters in the Nevada as of the 2022 primary, and of these 1,279,597 (slightly more
8
than 70 percent) resided in Clark County.
The rightmost plot in Panel (a) of Figure 1 shows that Election Day voters in Clark
County precincts supported Joseph Lombardo, the winning 2022 Republican gubernatorial
candidate, at slightly lower rates than did corresponding early voters. Note that the majority
of the points in the rightmost plot in this panel fall above the plot’s 45-degree line. In the 2014
and 2018 Republican gubernatorial primaries, no such relationship exists for the winners of
these two contests, Brian Sandoval and Adam Laxalt, respectively.
With respect to Panel (b), the cloud of points in the 2022 primary plot shows that mail
ballot voters supported Lombardo at greater rates than did corresponding early voters. This
occurred in the 2014 Republican gubernatorial primary as well, in which mail ballot voters
supported Sandoval at greater rates than did corresponding early voters. There are fewer
precincts pictured in the leftmost plot of Panel (b) because, prior to the passage of the
aforementioned Assembly Bill 321, mail ballot voting was not nearly as common in Nevada
as it is now. Sandoval won his primary overwhelmingly (the points in the two leftmost plots
7
“Governor Sisolak signs groundbreaking legislation to expand voting access in Nevada, increase education
funding,” Nevada Governor Steve Sisolak, June 2, 2021, available at https://gov.nv.gov/News/Press/2021/
Gov signs groundbreaking legislation expand voting access/ (last accessed September 19, 2022).
8
Voter registration statistics are drawn from the website of the Nevada Secretary of State, available at
https://silverstateelection.nv.gov/vote-turnout (last accessed September 19, 2022).
11
2014 Primary 2018 Primary 2022 Primary
100
75
Early vote
50
25
75
Mail
50
25
Note: includes only precincts with at least 25 voters for each voting method.
Figure 1: Election Day and early voting precinct vote shares of winning candidates in the
2014, 2018, and 2022 Republican gubernatorial primaries in Clark County
of Panels (a) and (b) are up and to the right) but a mail versus Election Day difference in
candidate support rates is evident in the 2014 Republican gubernatorial primary regardless.
There are no credible claims that this primary was subject to vote manipulation.
One could in principle argue that some form of vote manipulation affected the 2022
Republican gubernatorial primary by driving down the Election Day vote shares of the
primary’s winner, who just happened to be Joseph Lombardo. However, a quick look at the
12
100 100
75 75
Early vote
Mail
50 50
25 25
0 0
0 25 50 75 100 0 25 50 75 100
Election Day Election Day
,
(a) Election Day versus early voting (R) (b) Election Day versus mail (R)
100 100
75 75
Early vote
50 Mail 50
25 25
0 0
0 25 50 75 100 0 25 50 75 100
Election Day Election Day
,
(c) Election Day versus early voting (D) (d) Election Day versus mail (D)
Note: includes only precincts with at least 25 voters for each voting method; “D” denotes
the Democratic primary and “R,” the Republican primary.
Figure 2: Election Day, early voting, and mail ballot precinct vote shares of winning
candidates in 2022 United States Senate primaries in Clark County
2022 United States Republican and Democratic United States Senate primaries dispels this
conjecture. All three of these primaries appeared on 2022 Nevada primary ballots, and vote
shares from winning United States Senate candidates are depicted in Figure 2.
From Panels (a) and (b) of this figure, it is evident that, within Clark County precincts,
the winner of the Republican Senate primary did better with Election Day voters than
13
with early and mail ballot voters, the opposite of what occurred in the 2022 Republican
gubernatorial primary. Regarding the Democratic United States Senate primary, for which
there were no known credible allegations of vote manipulation, the winning candidate of
this contest did worse with Election Day voters than with early and mail ballot voters.
Overall, the lack of consistency across the four panels of Figure 2 is not consistent with vote
manipulation in 2022 Nevada primaries that, say, systematically diminished vote shares of
winning candidates.
Differing candidate vote shares by voting method are also apparent in the three most
recent presidential elections in Nevada. For Clark County precincts, Figure 3 plots Election
Day and early voting vote percentages for Barack Obama, Hillary Clinton, and Joe Biden, the
winning (Nevada) presidential candidates in the 2012, 2016, and 2020 presidential elections,
respectively. Within Clark County precincts, Obama in 2012 did marginally better among
early voters than Election Day voters, Clinton in 2016 clearly had more early voter support
than Election Day support, and there is little evidence in Figure 3 of a relationship between
Election Day and early voting support for Biden in 2020.
75 75 75
Early vote
Early vote
Early vote
50 50 50
25 25 25
0 0 0
0 25 50 75 100 0 25 50 75 100 0 25 50 75 100
Election Day Election Day Election Day
Note: includes only precincts with at least 50 voters for each voting method.
Figure 3: Election Day and early voting precinct vote shares of winning presidential
candidates in the 2012, 2016, and 2020 general elections in Clark County
Continuing with our analysis of recent presidential races in Nevada, candidate vote shares
14
from Election Day and mail balloting in 2012, 2016, and 2020 are displayed in Figure 4. In
2012, there is some evidence of a relationship between Obama’s mail ballot vote share and
his Election Day vote share; namely, Obama performed worse among mail ballot voters than
Election Day voters. In both the 2016 and 2020 presidential contests, Clinton and Biden,
respectively, clearly performed better among mail ballots than on Election Day ballots, the
latter election arguably exemplifying the jump in mail ballot voting that occurred particularly
among Democratic voters in 2020 (Yoder et al. 2021).
75 75 75
Mail
Mail
50 50 50
25 25 25
0 0 0
0 25 50 75 100 0 25 50 75 100 0 25 50 75 100
Election Day Election Day Election Day
Note: includes only precincts with at least 50 voters for each voting method.
Figure 4: Election Day and mail precinct vote shares of winning presidential candidates in
the 2012, 2016, and 2020 general elections in Clark County
In summary, a cursory look at only a few elections in Nevada turns up examples of races in
which vote shares of winning candidates were not equal across methods of voting. This is not
surprising given literature on how voters sort themselves into Election Day, early, and vote
by mail voting. Thus, the Gilbert contest’s claim — that a “fair election” features similar
candidate vote shares across voting methods, something “consistent with Clark County’s
Historical Election Results in all years prior to 2020, both in the Primaries and the General
Elections” (p. 8) — is false.
15
3.2 Nevada election results are not perfectly predictable
The Gilbert contest alleges that it is possible to predict to a suspicious degree of accuracy
Lombardo’s mail ballot vote total in any given Nevada precinct using only Gilbert’s mail
ballot vote total and Lombardo’s and Gilbert’s early voting totals. The contest contends
that this is evidence of vote manipulation, the ostensible logic being the idea that, if a feature
of Lombardo’s vote share is predictable, it must have been illicitly fixed prior to an election.
In this section, we demonstrate that there is nothing surprising or suspicious about what
the Gilbert contest calls the predictability of Lombardo’s vote shares in the 2022 Republican
gubernatorial primary. In fact, the contest’s supposed predictions have little to do with
actual election results. Instead, the contest proposes algebraic transformations of observed
vote totals that have nothing to do with the contest’s stated claims and yet mechanically
create the impression of a predictable election result.
Allegations of suspicious predictability appear throughout the Gilbert contest, but the
contest’s exact claim about this subject is difficult to pin down. For example, Daugherity
argues that vote totals for Gilbert and Lombardo disaggregated by mail-in ballots, Election
Day ballots, and early ballots are so “tightly dependent” that “the exact number of mail-in
votes for Lombardo” can be “exactly predicted. . . [in] every single precinct in [Clark] county”
(Appendix F, p. 4, ¶ 13). The body of the contest makes a related claim, positing that a
“formula fits all precincts in [Nevada counties] without any variation to such formula” (p. 9,
¶ 32, 5th bullet point), allowing one to “solve [for a] candidate’s aggregate percentage share
of the ballots in each [Nevada] precinct with no knowledge of Mail-in to Election Day Votes”
(p. 10, ¶ 32). Solomon and Allen make similar claims (Solomon’s Appendix A, p. 2, and
Allen’s Appendix D, p. 2, ¶ 5).
The Gilbert contest’s fundamental predictability argument is that information about
mail-in votes from one gubernatorial candidate (here, Gilbert) was used across Nevada
precincts — ostensibly by an unspecified individual or set of individuals — to determine
the number of votes Lombardo would need to win Nevada’s 2022 Republican gubernatorial
16
primary. Then, a formula was implemented that manipulated cast ballots to guarantee
that Lombardo received the requisite number of votes, this resulting in a functional — and
thus highly predictable — relationship between total votes cast for varying candidates.
While as evidenced above there are a number of ways across the Gilbert contest’s many
components in which predictability claims are adduced, here we analyze the prediction
argument from Daugherity’s Appendix F, this argument being the most coherent of the
prediction claims in the contest.
Daugherity divides 2022 Nevada Republican gubernatorial primary vote totals into two
categories: votes for Lombardo (the primary winner, as we have already noted) and
everyone else in the primary, labeled by Daugherity as “Gilbert et al.” Daugherity then
focuses on four variables for each precinct, defining them as follows (p. 3, ¶ 10):
Referring to his variables a - d, Daugherity states that, “Clearly these numbers should be
independent. . . [and that] knowing some of the numbers should not allow exactly predicting
the other numbers” (p. 3, ¶ 11). This claim is the basis of his predictability argument.
3.2.2 Lombardo’s and Gilbert et al.’s mail-in vote totals are linearly dependent
Daugherity’s claim about independence of the variables a - d is vague because it does not
specify if Daugherity believes these variables are mutually independent or pairwise
independent. The latter is weaker than the former, and thus we adopt this form of
17
independence here.
In any Nevada precinct, consider Daugherity’s claim that a (Lombardo’s mail-in vote
total) “should be independent” of b (Gilbert et al.’s mail-in vote total). This (pairwise)
claim must be false because by definition a and b sum to the total number of mail-in votes
cast in the precinct. Given mail ballot turnout, if in any Nevada precinct a increases by one,
then b must decrease by one. In other words, a and b are linearly dependent conditional on
9
mail-in voter turnout.
A parallel argument applies to Daugherity’s c and d, which are linearly dependent
conditional on early voting turnout. Therefore, Daugherity’s claim in the Gilbert contest
that his variables a - d “should be [pairwise] independent” is false.
Linear dependence aside, Daugherity’s claim about the independence of a - d (that there
should not be any) fails given known patterns of voter behavior in the United States. Voters
across the country evaluate residential locations in a way that is correlated with political
party affiliation (Gimpel and Hui 2015), and partisans sort themselves geographically (Brown
and Enos 2021). Therefore, one should expect there to be locations in Nevada that had
many supporters of Lombardo and other locations with many Lombardo opponents. This is
certainly true empirically at the county level, with Lombardo having received a larger vote
share in Clark County than in Nevada’s more rural Washoe County.
It follows from this logic that, in areas of Nevada where a is high, we would expect c to
be high as well. This does not mean that the correlation between a and c (and similarly
between b and d) should be expected to be literally one. Rather, if some areas of Nevada
were relatively pro-Lombardo (anti-Lombardo) in mid-2022, these areas will have relatively
9
For expositional simplicity, here we ignore mail-in ballots cast without valid votes in Nevada’s 2022
Republican gubernatorial primary as well as rejected mail ballots. If we were to include ballots like these,
we would instead posit that a and b must sum to the total number of non-rejected mail ballots cast in a
precinct with valid gubernatorial votes.
18
large values of a and c (b and d). This logic, gleaned from empirical literature on American
voters, renders as false Daugherity’s assertion that a, b, c, and d “should be independent.”
3.2.4 Daugherity’s claim that Lombardo’s mail-in vote totals can be predicted
with certainty is empirically false
Voter sorting notwithstanding, Daugherity’s claim that Lombardo’s mail in vote total (a)
“can be exactly predicted from b, c, and d. . . [in] “every single precinct in [Clark county]”
(p. 4, ¶ 13) fails because it is simply not true. To assess this claim, we regress Lombardo’s
mail-in vote total (a) on variables b (Gilbert et al.’s mail in vote total), c (Lombardo’s early
in person vote total), and d (Gilbert et al.’s early in person vote total). If Daugherity’s
claim were correct, then a could be predicted from b, c, and d, with essentially no residual
10
variance.
Dependent variable:
a
∗∗∗
b 0.886
(0.027)
∗∗∗
c 0.602
(0.053)
∗∗∗
d −0.347
(0.045)
Constant 0.011
(0.834)
Observations 647
2
R 0.863
Residual Std. Error 13.102 (df = 643)
∗ ∗∗ ∗∗∗
Note: p<0.1; p<0.05; p<0.01
Results from our regression are in Table 1, and they show that Daugherity’s a cannot
10
The regression summarized in Table 1 includes only Clark County precincts that have a > 0, b > 0,
c > 0, and d > 0. This is because Clark County’s official precinct-level election results suppress reported
vote totals (i.e., variables like a - d, among others) when these totals are very small. When, say, a is not
reported for a given precinct, we cannot determine its value.
19
be predicted by the variables b, c, and d. In fact, our regression has a residual standard
error of approximately 13 mail ballot votes per precinct. To put this in context, on average
Lombardo received roughly 38 mail-in votes in each Clark County precinct, implying that
a residual standard error of 13 is substantial and far from being consistent with an exact
prediction.
To ensure that Table 1’s lack of predictability in a is not due to a functional form
limitation in the table’s linear regression, we also predicted Lombardo’s mail-in vote total
a based on b, c, and d using (1) a random forest regression and (2) a generalized additive
11
model (GAM). Results from these two model are in Table 2.
Table 2: Root mean squared errors for linear regression, random forest, and generalized
additive models (GAM), all of which seek to predict a based on b - d.
The leftmost column of Table 2 repeats the value of 13.102, the residual standard error
from our aforementioned linear regression model. In the middle column of the table, we
present the root mean squared error from a random forest regression, which is a flexible
machine learning tool that does not make explicit functional form assumptions about the
relationship between a dependent variable (here a) and a set of explanatory variables (b - d)
(Hastie et al. 2009). Using a random forest, we find a root mean square error of approximately
12
15.8 votes — a non-trivial error given typical Lombardo vote totals per precinct. Finally,
in the rightmost column of Table 2, we present the results from our GAM, which is a tool
that assumes the relationship between covariates (here, b - d) and the outcome variable (a)
is additive but beyond this makes no assumptions about the functional form that describes
the relationship between covariates and outcome. Using a GAM, we find a root mean square
11
We fit our GAM with smooth terms for variables b, c, and d, with cross validation to determine model
complexity as implemented in the R package MGCV.
12
While a more flexible method, a random forest regression can perform worse that a linear model if the
relationship being modeled is approximately linear and the random forest regression uses information to
estimate the functional form assumption from a linear regression.
20
error of 12.3 votes — again, a substantial error.
Our root mean squared error results in Table 2 reflect in-sample calculations, meaning
that we used the same data to estimates our models and then to draw conclusions. This biases
our estimates of root mean squared errors towards zero, which is in favor of Daugherity’s
claim that a can be predicted by b - d.
Thus, according to standard statistical methods of prediction, there is not an exact
relationship between Lombardo’s mail-in vote total and the other vote totals. To the point,
Daugherity’s claim that his variable a can be predicted by his variables b - d is false.
The Gilbert contest, and Daugherity in particular in his Appendix F, argues that an exact
prediction of a can be obtained based on a set of three non-linear transformations of a, b, c,
and d proposed by Solomon. In line with Solomon’s argument in his Appendix A (p. 14),
Daugherity describes three new variables g, h, and α that are functions of his original a, b, c,
and d (Appendix F, p. 4, ¶ 15). These variables are defined as follows:
d
g = (1)
a+d
b
h = (2)
b+c
b+d
α = (3)
a+b+c+d
With these new variables, Daugherity estimates the following precinct-level regression
(Appendix F, p. 5, ¶ 18):
g = B0 + B1 × α + B2 × h + ei
21
Based on the definitions of g, h, and α, this is equivalent to the following:
d b+d b
= B0 + B1 × + B2 × + ei (4)
a+d a+b+c+d b+c
Reinforcing the point made above, the variable a is on both the left and right hand sides of
the regression in Equation 4, making it obvious that the regression cannot be used to predict
a, or any function of a for that matter.
Daugherity’s Appendix F provides the following estimates (with the level of precision
presented exactly as on ¶ 18 of the appendix):
Of the values of g, h, and α that produced the regression estimates in Equation 5, Daugherity
states that they exemplify “improper dependence [which] confirms that the election results
in the June 14, 2022, Republican gubernatorial primary in Clark County, Nevada, were
artificially contrived” (Appendix F, ¶ 17). And in a revised appendix to the Gilbert contest,
2
Allen claims to estimate the same regression and reports obtaining an R of 0.96, arguing
like Daugherity that this “is extraordinarily high” (¶ 20).
The ostensible intuition on which Daugherity and Allen are drawing here is that a high
2
value of R in Equation 5 implies that the 2022 Republican gubernatorial primary must have
been “fixed” prior to the election. In other words, if Nevada’s 2022 Republican gubernatorial
2
primary produced values of g that when regressed against h and α return a high R value,
then it must be the case, according to Daugherity, that the election suffered from vote
manipulation.
Putting aside the fact that a - d are not independent in the first place, the fact that a
regression with (1) a function of a on its left-hand side and (2) a different function of a on
2
its right-hand side has a high value of R implies nothing at all.
Moreover, the three non-linear transformations of a - d in Equations 1 - 3 that Daugherity
22
used to generate g, h, and α induce a strong relationship between these three latter variables,
even in instances where there is no malfeasance at all associated with the variables a, b, c,
2
and d. That is, a high R value in the regression in Equation 4 is a feature of the regression
model and thus implies nothing about the underlying data.
We provide details on this claim in Appendix A, which presents three sets of simulations
based on Clark County precincts from the 2022 Republican gubernatorial primary. In each
such set, we randomly draw the variables a - d in a non-illicit way, i.e., in a way that does
not presuppose that there was any vote manipulation. After drawing a - d in what can be
called a clean way, we calculate simulated values of g, h, and α and then estimate Equation
2
4. We find high values of R , indicating that what Daugherity and Allen claim is evidence
of vote manipulation is in fact evidence of nothing.
Daugherity appears unaware that the underlying regression model of Equation 4 whose
2
estimates appear in Equation 5 will generate a high R regardless of the presence of vote
2
manipulation. Building on his (erroneous) logic that the high R does indeed show
malfeasance, Daugherity goes on to claim that, if one knew the values of α, b, c, and d,
then one could predict the value of a (Appendix F, ¶¶ 20-21). Daugherity explains this
using a hypothetical example, assuming that α = 0.463855422 in a given Clark County
precinct. Then, given values of b = 95, c = 45, and d = 133, it follows that a = 133.
This does not represent, however, a “prediction” of a. Since α is a function of a, as made
clear in Equation 3, it is not possible to know the value of α in a Clark County precinct (or in
any precinct for that matter) without first knowing the value of a. To state the obvious and
place this issue in terms of the alleged vote manipulation described in the Gilbert contest,
if the transformations that underlie g, h, and α are intended to capture how some nefarious
actor or actors were to have manipulated Lombardo’s mail in vote total (a), the way the
manipulation was carried out, as exemplified by the transformations, cannot have depended
upon the number of mail-in votes Lombardo received (a).
Despite Daugherity’s claim that he can make an exact prediction of Lombardo’s mail-
23
in vote total a in each Nevada precinct based on a set of four other variables (α and b
- d), nothing in fact in Daugherity’s report is actually a prediction. Rather, in his work
Daugherity merely demonstrates a series of algebraic manipulations that move from one set
of variables (a - d) to another (g, h, and α). As these manipulations can only be carried
out when the variables a, b, c, and d are all known in the first place, all that Daugherity
shows in his contribution to the Gilbert contest is that, if one assumes a given variable has
a particular value in an algebraic equation, then one will find that variable has the value
assumed. By way of analogy, Daugherity’s analysis is equivalent to someone claiming an
ability to predict the price of a house based on knowing (1) the house’s square footage and
(2) its price per square foot.
The Gilbert contest was a legal challenge to a 2022 Republican primary election in Nevada.
Its claims — that fair elections have equal candidate vote shares across voting methods,
that mail-in and early voting vote shares of candidates are independent, and that a form of
predictability indicates the presence of malfeasance in the primary, among others — range
from unfounded to simply false. Taken together, the claims in the Gilbert contest fail due to
manifold errors of statistics, logic, and misrepresentation of election history. They also fail
because their supposedly surprising predictive results are merely the consequence of using
transformations of variables to predict themselves.
Nonetheless, these claims were the basis of a challenge to an election winner. They
exemplify how an expansive theory of vote manipulation, one not grounded in evidence, can
lead to a real legal dispute.
The same types of errors that affect the Gilbert contest manifest themselves in our next
set of claims about illicit vote manipulation, to which we now turn.
24
4 Douglas Frank’s many theories of election
malfeasance
After the 2020 election, Douglas Frank, a Ph.D. chemist and high school math teacher, rose
to prominence asserting he had clear evidence of malfeasance in the 2020 election (Seidman
2021). “Just about every county in the country was hacked,” he said of this election
(Murray and Simon 2022). Subsequently, Frank has toured the country, meeting with
elections officials during the day and citizen groups during the night, to spread his vote
manipulation theory (Gardner et al. 2021; Halpern 2022). Frank alleges an expansive vote
manipulation conspiracy that in principle encompasses every election in the United States,
in every state and in every county. If true, Frank’s theory would call into question the
position of essentially all elected officials in the country.
While Frank’s story sometimes varies, his core assertions about vote manipulation rest
on three steps, which we have deduced from videos of his public presentations. First, Frank
argues that voter rolls are inflated with “phantom voters” (Frank 2021b). Per Frank,
phantom voters are registered voters who are no longer eligible to vote in a particular
district or potentially were never eligible to begin with. According to Frank, these
individuals could have lost eligibility because they moved, because they died, or because
they never existed at all. Phantom voters are important, Frank contends, because they
constitute the voters who can be inserted into election results to obtain a desired outcome.
Second, Frank alleges that there exists in U.S. elections a “key” that is determined at the
state level and then deployed across a state’s counties by an unspecified group of conspirators
(Frank 2021c). Specifically, Frank contends that such a state-level key is based on state-level
turnout rates for registered voters grouped by age, 18 to 100. According to Frank, for a
given election a state’s key is used to determine, prior to the election, the reported number
of voters from each age group in each county who turn out to vote. In a national election,
Frank alleges that each state in the country uses a different key.
Third and finally, Frank argues that, before final election results are reported, an
25
algorithm at the county-level administered by unnamed conspirators compares the number
of ballots cast in a county by voters of age a with the prescribed turnout rate for age a
based on a state-level key that was determined earlier. The difference between prescribed
turnout and actual turnout is then made up by inserting votes from a reservoir of
“phantom voters.”
While Frank alleges a coordinated nationwide effort to determine statewide voter turnout
rates, to the best of our knowledge he has never identified the specific perpetrators of the vote
manipulation scheme that he claims to have identified. Nor has he identified the organization
(or in principle organizations) that have the capacity to conduct this scheme. Frank has made
vague allusions about funding from Meta Chief Executive Officer Mark Zuckerberg (Frank
2022a), but to the best of our knowledge he has never specifically explained Zuckerberg’s
role in an alleged vote manipulation conspiracy.
Moreover, to the best of our knowledge, Frank has never presented direct evidence of the
conspiracy that is summarized in the three steps above. Instead, he argues that the existence
of the conspiracy can be inferred from indirect evidence obtained from official election returns
and statewide voter registration list. A statewide voter registration list is a list of registered
voters in a given state that contains voter characteristics like name, address, age, and the
historical elections in which voters have turned out. Statewide voter registration lists are
13
often called statewide voter files.
For instance, Frank argues in favor of the existence of “phantom voters” based on
well-documented issues associated with voter file maintenance. In particular, he asserts
that differences between the total number of reported voters in a given election based on a
statewide voter file and actual turnout in the election reported by state officials is evidence
of malfeasance. It is not, however. In Appendix B, we explain that what Frank considers
troubling reflects the fact that records of registered voters are regularly deleted from
statewide voter files when voters move, die, or cancel their voter registrations.
13
States vary in the voter characteristics that are recorded in their voter files. Moreover, some states make
their voter files public while others do not. See National Conference of State Legislatures (2022).
26
Voter file maintenance issues aside, we focus here on addressing Frank’s statistical claim
that he has developed a method to diagnose whether an election suffers from vote
manipulation. Contrary to Frank’s claims, we show that his proposed test for the presence
of manipulated votes will regularly diagnose fair elections as suffering from manipulation.
This is because his test effectively correlates a variable with itself, a phenomenon we also
saw in our earlier discussion of the Gilbert contest.
In a given election, Frank’s test for vote manipulation is based on calculating the correlation
between a predicted number of voters in an age group in a county (where predictions are based
on a state-level key that Frank asserts is part of a vote manipulation conspiracy) and the
observed number of voters in that age group within a county. Operating across age groups,
Frank calculates one correlation per county and reports that his correlations are often close
to one for practically all counties in a state, which he asserts is evidence of malfeasance. The
ostensible logic here is that a high correlation between observed voter turnout and turnout
as predicted by a vote manipulation conspiracy implies that the conspiracy was at work.
Here we show, however, that Frank’s test fails as a tool to detect vote manipulation and in
fact will return a correlation between predicted and observed turnout numbers close to one
even if the county being scrutinized suffered from no vote manipulation at all.
Turning now to details, Frank’s test operates as follows. In a state s and for voters
of age a ∈ A = {18, . . . , 100}, Frank calculates the turnout rate tsa , usually basing such
a calculation on observed turnout nsa among voters of age a in state s and dividing this
number by the number of registered voters vsa of age a in state s. As a simple matter of
algebra, it is useful to note that nsa = tsa × vsa , i.e., that the total number of voters of age
a in state s is the product of a turnout rate and a count of registered voters. To calculate
age-based turnout rates within each state, Frank uses statewide voter files, which as we have
noted often include voter age.
27
For the age groups a ∈ A in state s, Frank regresses age-based turnout tsa rates on a
sixth order polynomial of underlying age and a constant:
2 3 4 5 6
tsa = β0 + β1 a + β2 a + β3 a + β4 a + β5 a + β6 a + ϵsa (6)
For a state s, the estimated coefficients from the regression in Equation 6 yield fitted values
t̂sa for a ∈ A. These fitted values are predicted turnout rates, and Frank alleges that each
county i in state s will have the same predicted turnout rate for registered voters of age a
(we discuss this allegation shortly and for the moment take it at face value). Based on this,
Frank calculates a predicted number of voters from each age a in each county i that is part of
state s. To do this, for age group a Frank multiplies the statewide predicted turnout rate t̂sa
with the number vsia of voters in age group a, county i, and state s. That is, the predicted
̂ sia = vsia × t̂sa .
number of voters from age group a in county i in state s is n
Finally, Frank calculates the correlation within each county i between the predicted
̂ sia of voters in each age group a and the observed number of voters nsia in that
number n
group:
Cor (̂
nsia , nsia ) = Cor (t̂sa × vsia , tsia × vsia ) (7)
where tsia is the observed turnout rate for voters in county i of state s who are a years old.
Frank claims that, if this correlation is large, then the state-level “key” determined actual
election turnout and, accordingly, that vote manipulation occurred (Frank 2021a). The
underlying premise behind this ostensible logic is that county-level election results should
not be so regular as to have high correlations between what Frank calls predicted turnout
and observed turnout. Frank claims — by way of analogies and simple assertions rather
than via any theoretically driven rationale — that high predictability as highlighted by
high correlations in Equation 7 is “unnatural” and suspicious. For example, Frank regularly
claims that correlations above 0.7 are rare when analyzing human behavior and therefore a
28
correlation close to one is unnatural (for example, Frank 2021a). This is, of course, untrue
— there is no general cutoff for a surprising correlation within any field of study. Rather
than reasoning by analogy and simple assertion, a rigorous evaluation of Frank’s test would
examine whether correlations close to one in Equation 7 occur only in elections where votes
have been manipulated in some fashion. We now provide such an evaluation and show that
a correlation close to one should be expected in almost all instances based on the calculation
in Equation 7, whether there is vote manipulation or not.
A fundamental feature of Frank’s test for vote manipulation is that it calls for calculating
the correlation in Equation 7 between the predicted and actual number of voters in an age
group at the county level. This will mechanically result in correlations that are close to one.
This occurs because Frank’s test essentially “discovers” the regularities that (1) age groups
with more registered voters tend to have more individuals turning out to vote while (2) age
groups with fewer registered voters conversely tend to have fewer people turning out to vote.
Because in any county there is considerable variation in the number of voters in any age
group, and relatively less variation in the rate at which individuals in an age group turn
out to vote (these rates being constrained between zero and one, by definition), observed
numbers of voters (nsia ) and “predicted” numbers of voters (̂
nsia ) move together relatively
closely, which results in a correlation close to one in Equation 7. In fact, one way to view
Frank’s test encapsulated in this equation is that it calls for effectively correlating a variable
with itself: the number of registered voters vsia of age a in county i of state s dominates
both the left and the right input of the correlation.
29
4.3 Why Frank’s test for vote manipulation is meaningless
To see this self-correlation, let us revisit Equation 7, paying particular attention to the fact
̂ sia and nsia . Accordingly, and fixing state s and
that Frank’s test involves a correlation of n
county i, notice that the number vsia of registered voters in age group a appears in both
the left and the right input of the desired correlation. The result of this is that variation in
the number of registered voters will cause both predicted (left side of the correlation) and
observed numbers of voters (right hand side) to move together. Ultimately, this results in
correlations close to one.
A simple simulation suffices to demonstrate the mechanical correlation inflation that
occurs when correlating numbers of predicted voters and numbers of observed voters in age
group a rather than predicted turnout rates and observed turnout rates for these individuals.
The simulations we construct are free of vote manipulation. Therefore, to the extent that
they have high correlations based on Equation 7 — particularly as we increase the variance
in the number of individuals in each age group — it follows that Frank’s test fails as a
diagnostic of vote manipulation.
To construct our simulations, we suppose like Frank that there are 83 distinct ages in a
hypothetical county within a single state, individuals ranging from 18 to 100 years old. To
show that variation across age group size drives the correlation at the root of Frank’s test,
we further suppose that 82 of these age groups have 100 individuals — so, 100 18 year olds,
100 19 year olds, and so forth. In the simulations, however, we vary the number of people in
the final age group, which allows us to observe how the correlation calculated in Equation
7 changes as the variation in the number of people in each age group changes. Specifically,
across simulations we suppose that the number of registered voters in the target age group
varies from 100 — the same size as other age groups — to 1000, i.e., ten times the number
of individuals in the largest age group relative to other age groups.
Conditional on numbers of voters, we simulate turnout in an election in a way that is
independent of age. In each simulated election, we suppose that the turnout rate of each
30
age group a is 40 percent plus random error that is independent and identically distributed
Gaussian noise. Formally, our simulations posit that voter the turnout rate for age group a
is 0.4 + γa where γa ∼ Normal(0, 0.025). To assess Frank’s test, within each simulation we
fit a sixth order polynomial with Equation 6 to obtain a “key” and then, within a simulated
county, use this to calculate a predicted number of voters in each age group.
Table 3 presents the results of our simulations. The left-most column in the table
(“Largest age group size”) lists the size of the largest age group in each of the simulated
elections. The middle column (”Number of voters”) reports average correlations between
Frank-predicted turnout numbers and observed turnout numbers. And, the right-most
column of the table (“Turnout rates”) reports reports average correlations between
predicted turnout rates and observed turnout rates.
Average correlation
Largest age group size Number of voters Turnout rates
100 0.26 0.26
200 0.68 0.26
300 0.87 0.26
400 0.94 0.26
500 0.96 0.26
600 0.98 0.26
700 0.98 0.26
800 0.99 0.26
900 0.99 0.26
1000 0.99 0.26
The key implication of Table 3 is straightforward: as the size of the largest voter age
group increases relative to other age groups, the correlation between numbers of predicted
voters based on Frank’s “key” model and numbers of actual voters approaches one. This is
because, when there is more variation in the number of individuals in a set of age groups,
predicted and observed numbers of voters move closely together, which is merely the result
of the correlation in Equation 7 uncovering the unsurprising regularity that larger age groups
tend to have more voters. In contrast, and as made clear in the rightmost column in Table
31
3, if one focuses on turnout rates, which control for age group size, the correlation between
predicted and observed turnout holds steady at 0.26. In short, Table 3 demonstrates that the
supposedly unnatural correlations from Frank’s test for vote manipulation can occur merely
because of variation in the number of individuals across different age groups.
We also see a mechanical increase in correlation when we analyze actual election data.
Because Frank has not responded by October 24, 2022 to our request for data and replication
code for his nationwide analysis (see Appendix C), we independently evaluated his claims
about the utility of his proposed test for vote manipulation. To do this, we used voter file
data from L2 and calculated turnout rates for registered voters for each age from 18-100
14
for all states where we observed a birth date for more than 90% of registrants. Our L2
voter file data derive from snapshots of states’ voter files; since states do not synchronize
updates to their respective voter files, pull dates vary across states between December 2020
and May 2021, as described in Appendix E. With turnout rates calculated at the state level
for each age, we followed Frank’s procedure and Equation 6 and regressed these rates on
a six-degree polynomial to create predicted turnout rate for each age. State by state, we
then projected predicted turnout rates to counties. Continuing to follow Frank’s procedures,
we calculated the correlation within each county between observed numbers of voters and
predicted numbers of voters by age. For a given age, predicted numbers of voters were
obtained by multiplying the number of registrants of that age by the predicted turnout rate
for that age where prediction is based on Frank’s key procedure.
In Figure 5, we present the results of applying Frank’s analysis to the 42 states where
we have adequate age data. In this figure, the vertical axis describes statewide average
correlation coefficients between predicted and observed turnout numbers (i.e., average values
of Cor (̂
nsia , nsia )). To generate these numbers, in each county contained in our set of 42
states we calculate a correlation across age groups between predicted and actual turnout
numbers. Then, for each state we average these county-level correlations to create a single
14
Excluded states were Alaska, the District of Columbia, Hawaii, Maryland, Mississippi, New Hampshire,
North Dakota, Wisconsin, and Wyoming.
32
statewide correlation. Continuing, Figure 5’s horizontal axis describes statewide average
correlation coefficients between predicted and observed turnout rates (i.e., average values of
Cor (t̂sa , tsia ) where t̂sa is the predicted turnout rate for voters of age a in county i of state
s, where t̂sa is not indexed by i because Frank makes a single turnout rate prediction for
age group a in the entire state). Parallel to the procedure described above, in each county
in each state we calculate the correlation between the predicted and actual turnout rates.
Then, in each state we take the average of the county-level correlations to obtain a single
statewide average correlation between predicted and actual turnout rates. Figure 5 includes
a dashed 45-degree line, and all 42 points in the figure lie above it, indicating that, in all 42
states we study, the average correlation coefficients between observed and predicted turnout
numbers is greater than the average correlation coefficients between observed and predicted
turnout rates.
Figure 5 shows that in every state we find a higher average correlation between predicted
and actual turnout numbers then we do between the predicted and actual turnout rates.
In fact, turnout number correlations average 0.991 across the U.S. counties in our data set,
with an interquartile range from 0.9905 to 0.9978. In contrast, rate correlations average 0.825
across counties in our data set, with an interquartile range of 0.758 to 0.922. Taken together,
this shows correlating turnout numbers artificially inflates the reported correlations.
In our discussion of Frank’s theory of vote manipulation, we noted that the theory implies
that, within states, voter turnout rates by age would be effectively constant across counties.
This is most decidedly not the case, which undermines Frank’s argument that turnout rates
are fixed by a “key.”.
Figure 6 displays by state voter turnout rates (vertical axis) by age (horizontal axis) for
the 2020 general election. Each point in this figure is a turnout rate for a particular age
33
ID LA TN MO MA VTNCCAFL NJ CT ME RI
1.0 MN
UT KY
MT NV CO WV IN AL MI SC DE
NE NY
KS GA NM OR OH AZ
SD TX
IA AR PA OK
VA IL WA
Average correlation in 0.9
turnout counts
0.8
0.7
0.6
Figure 5: High correlations between actual and predicted ballots in counties is found in each
state.
group in a county, and what Frank would call the “key” for each state is approximated by
15
the red line.
Consider Texas, which had the second largest registered voter pool in the 2020 general
election. Turnout rates of the youngest Texans are shown on the left side of Figure 6’s Texas
panel, and turnout rates of the oldest Texans are shown on right side of the panel. The
red curve in the Texas panel is roughly concave down, indicating that, in the 2020 general
election, younger and older Texans had lower turnout rates than middle age Texans.
Within each state and conditional on age, Figure 6 highlights considerable county-to-
county variation in turnout rates across the 42 states included in our analysis. Continuing
to examine Texas, there are 254 counties in this state. Focusing on the turnout rates of
15
The red lines in the figure are loess smoothers, with bandwidth determined by cross-validation.
34
Figure 6: Voter turnout rates and age across states in the 2020 general election
Note: states are ordered by number of registered voters as of the 2020 general election.
the youngest voters in the Texas (age of 18), county points vary and appear well below and
above the red curve. This is true as well of 19 year old voters and in fact of voters of all
ages. That is, voters of age a ∈ [18, 100] did not have identical turnout rates across Texas
counties. This is inconsistent with Frank’s vote manipulation theory.
At Mike Lindell’s “Moment of Truth” summit, held August 20-21, 2022, Frank claimed
to have evidence that vote counts in California had been manipulated in the 2020 general
election. Specifically, Frank asserted that precinct-level partisan turnout rates were constant
across counties in California, the sort of regularity that Frank takes to be evidence of vote
manipulation.
35
To make his case, Frank first focused on the relatively conservative Placer County,
located on the California-Nevada border near Lake Tahoe. Frank asserted that, “88% of
the registered Republicans in EVERY Placer County precinct voted...Exactly 88% of all
Republicans voted in that Precinct” (Frank 2022c). After asserting that this is implausible,
Frank presented a slide where he opined that, “we might suspect that someone was stuffing
ballots up to a target limit of 88%” (Frank 2022c). He went on to assert that the basic
pattern of across-precinct consistency in partisan turnout rates held for Democratic and
overall turnout in Placer County and for precinct-level partisan turnout in California
counties statewide claiming that, “this is all over the state, this is happening everywhere”
(Frank 2022c).
16
Frank’s analysis of California, however, is simply inaccurate. Once (1) Frank’s
statistical misunderstandings addressed and (2) we correctly disentangle turnout rates and
turnout numbers, there is no evidence of any surprising patterns in the 2020 general
election results in California.
The left-hand plot in Figure 7 replicates, as closely as possible, the plot Frank presented at
the aforementioned “Moment of Truth” summit in August 2022. To replicate Frank’s plot,
17
we used precinct-level data from the 2020 general election in California. The horizontal
axis in the left-hand plot of Figure 7 shows the number of Republican registered voters and
the vertical axis, the number of Republicans who turned out to vote. Pictured as well in the
plot is a blue least squares regression line, which summarizes the displayed points.
16
Frank claims that there were 35 precincts in Placer County in the 2020 general election. This is incorrect:
there were 127 voting precincts in the county. Further, the sizes of the precincts Frank described at the
“Moment of Truth Summit” are not at all reflected in official election results (Placer County Elections Office
2020). Finally, at the summit Frank displayed a map of Placer County divided into rectangular grids. We
cannot determine such a network of grids corresponds to precincts in Placer County, if it does at all.
17
Our data on Placer County and other counties in California was gathered from the California
Statewide Database, available at https://statewidedatabase.org, with the specific data obtained at https:
//statewidedatabase.org/d10/g20.html (last accessed October 21, 2022).
36
Figure 7: Republican voter turnout in Placer County in the 2020 general election
Note: shows only Placer County precincts with more than 100 registered voters.
The left-hand plot in Figure 7 and its blue regression line, which is sloped up, reveals
the fact that, in the 2020 general election, larger precincts in Placer County had both more
registered Republicans and more Republican voters. This is not surprising.
The figure’s regression line, however, cannot speak to Frank’s claim about equal precinct-
level turnout rates in Placer County. That said, in his comparable figure, Frank reports a
regression slope of 0.88 where the slope of the blue line in the left-hand plot of Figure 7 is
0.9. Armed with his slope, Frank asserts that every precinct in Placer County saw 88% of
Republicans voting in the 2020 election. This is demonstrably false, as made clear by the
right-hand plot in Figure 7, which presents precinct-level Republican turnout rates across
Placer county. This plots shows that there is considerable variation in these turnout rates
— with a precinct-level high of 98 percent of registered Republicans and a low of 78 percent.
We observe the same pattern for Democrats and overall turnout in Placer County, which we
37
show in Appendix D.2.
38
39
Figure 8: Republican Turnout Rates for the 24 California Counties with the most registered voters. Across the counties there
is considerable variation in the turnout rate of precincts
Note: counties order by numbers of registered voters in the 2020 general election.
2 2
vijp and recorded the R for county j in that iteration of the simulation Rj . We repeated
this simulation 1,000 times for each county.
2 2
From our simulation, we find a very high average R . The average R for Democrats,
Republicans, and for overall turnout is 0.992, 0.994, and 0.992, respectively. This high
2
average R occurs even though the turnout rate in each precinct is independently drawn at
random. In short, Frank’s supposed evidence of vote manipulation in California precincts
is not evidence of manipulation at all. Rather, it demonstrates that precincts with more
registered voters tend to have a more voters turnout to vote.
In a variety of videos and speeches made since 2020, Douglas Frank has articulated a theory
of vote manipulation that covers essentially the entire United States. If Frank’s theory
were correct, practically every contemporary election in the country has been marred by
malfeasance.
In his work, Frank proposes a test for vote manipulation. While Frank asserts that the test
is truly diagnostic, in reality it shows nothing. Rather, Frank’s test for voter manipulation
fails because it always classifies elections as suffering from vote manipulation due to its
construction. As we have explained, this is because Frank’s test confuses variation in the
number of individuals across age groups in counties and states with evidence of an unnatural
manipulation of ballots.
Not only is Frank’s test for vote manipulation worthless, but a key principle on which it
is based — that various precinct-level turnout rates are constant within and across countries
— is wrong. In short, Frank’s findings have no implications at all for the presence of vote
manipulation in the United States.
40
5 Refuting conspiracy theories about election
malfeasance
In this paper, we have considered two prominent exemplars of new, expansive theories of vote
manipulation. One such theory, which formed the foundation of a formal election contest
filed in Nevada, was articulated by 2022 Republican gubernatorial primary candidate Joey
Gilbert after he lost his race to Joe Lombardo. In court documents, Gilbert claimed that
an algorithm had shifted votes from him to Trump-endorsed Republican Lombardo and the
incumbent Nevada governor Steve Sisolak, a Democrat. A second expansive theory of election
malfeasance has been spread by Douglas Frank, who alleges that an algorithm determines
the results of all United States elections.
We have shown that both Gilbert’s and Frank’s theories about elections are baseless.
Gilbert’s contest about vote manipulation misrepresented prior election results, asserted
conclusions that fail to follow from their premises, and merely showed that it is possible to
“predict” a variable when that variable is known. Frank’s claims are similarly vacuous. His
statistical tests merely show the unsurprising fact that groups with more registered voters
have more people who turn out to vote.
The utility of our arguments lies in identifying and refuting false claims about election
malfeasance and offering refutations that might be helpful for local elections officials and
others. Even so, the environment in which baseless election claims are engaged is challenging
because, at least in the case of the Gilbert contest in Nevada, those alleging malfeasance
seem remarkably uninterested in acknowledging their refutation. For example, in a hearing to
dismiss the Gilbert contest, Gilbert’s lawyer asserted that, “Nobody from his side. . . has said
that the formulas are wrong or that they generate errors in the mathematical computation
of the purported mail in ballots” despite reading lengthy reports from two of the authors of
this paper (Rumble 2022, 18:30).
The only path to defeat baseless election claims is demonstrating that they are false
and publicly explaining why. This is the motivation for this paper, whose results should be
41
valuable to all those who care about the infrastructure of American elections. These people
include election officials, who have to engage baseless concerns from concerned citizens,
judges who have to weight the evidence of malfeasance in court cases, and citizens who remain
concerned about American elections but are still open to the value of scientific evidence.
References
Ansolabehere, Stephen and Nathaniel Persily (2008). Vote fraud in the eye of the beholder:
The role of public opinion in the challenge to voter identification requirements. Harvard Law
Berlinski, Nicolas , Margaret Doyle, Andrew M. Guess, Gabrielle Levy, Benjamin Lyons, Jacob M.
Montgomery, Brendan Nyhan, and Jason Reifler (2021). The effects of unsubstantiated claims
Brown, Emma and Jon Swaine (2022). FBI seizes phone from election denier Douglas Frank, a
Brown, Jacob R. and Ryan D. Enos (2021). The measurement of partisan sorting for 180 million
Cassidy, Christina A. (2021). Far too little vote fraud to tip election to Trump, AP
Cole, Brendan (2022, June 15). Jan. 6 Attendant Loses Nevada Primary by 11 Points, Blames
Corasaniti, Nick and Alexandra Berzon (2022, June 26). The strange tale of tina peters.
42
The New York Times. https://www.nytimes.com/2022/06/26/us/politics/tina-peters-election-
Cottrell, David , Michael C. Herron, and Sean J. Westwood (2018). An exploration of Donald
Trump’s allegations of massive voter fraud in the 2016 General Election. Electoral Studies 51,
123–142.
Danforth, John , Benjamin Ginsberg, Thomas B. Griffith, David Hoppe, J. Michael Luttig,
Michael W. McConnell, Theodore B. Olson, and Gordon H. Smith (2022). LOST, NOT STOLEN:
The Conservative Case that Trump Lost and Biden Won the 2020 Presidential Election.
https://lostnotstolen.org//wp-content/uploads/2022/07/Lost-Not-Stolen-The-Conservative-
Case-that-Trump-Lost-and-Biden-Won-the-2020-Presidential-Election-July-2022.pdf (last
Eggers, Andrew C. , Haritz Garro, and Justin Grimmer (2021). No evidence for systematic voter
fraud: A guide to statistical claims about the 2020 election. Proceedings of the National Academy
Frank, Douglas (2021a). Dr. frank’s presentation at washington election integrity public
hearing. https://rumble.com/vmjuoo-dr-franks-presentation-at-washington-election-integrity-
public-hearing.html.
phantom-voter.html.
at-election-committee-12-8-21.html.
Frank, Douglas (2022a). Dr. douglas frank presents election data. https://rumble.com/v1ckbar-
july-17-2022.html.
https://rumble.com/v1639mg-election-fraud-in-idaho-dr.-frank-presentation.html.
43
Frank, Douglas G (2022c, August 22). Dr. Douglas G. Frank Speaks at The Moment
Gardner, Amy , Emma Brown, and Josh Dawsey (2021). Inside the nonstop pressure
campaign by Trump allies to get election officials to revisit the 2020 vote. The Washington
Post. https://www.washingtonpost.com/politics/trump-election-officials-pressure-campaign/
2022).
Gardner, Amy and Patrick Marley (2022, September 11). Trump backers flood election offices with
Gentry, Dana (2022). Gubernatorial hopeful Joey Gilbert: No regrets about Jan. 6. Nevada
Current. https://www.nevadacurrent.com/2022/01/06/gubernatorial-hopeful-joey-gilbert-no-
Ghitza, Yair , Andrew Gelman, and Jonathan Auerbach (2022). The Great Society, Reagan’s
Gimpel, James G. and Iris S. Hui (2015). Seeking politically compatible neighbors? the role of
Goel, Sharad , Marc Meredith, Michael Morse, and David Rothschild (2020, May). One Person,
One Vote: Estimating the Prevalence of Double Voting in U.S. Presidential Elections. American
Gordon, Aaron (2021). ‘Expert Mathematician’ on Election Fraud Actually a Swing Set Installer,
44
Halpern, Sue (2022). The Election Official Who Tried to Prove ‘Stop the Steal’. The New
Yorker. https://www.newyorker.com/news/american-chronicles/the-election-official-who-tried-
Hansen, Claire (2022, October 14). As jan. 6 committee discloses new details, trump responds with
2022-10-14/as-jan-6-committee-discloses-new-details-trump-responds-with-recycled-lines (last
Hasen, Richard L. (2022). Record Election Litigation Rates in the 2020 Election: An Aberration
Hastie, Trevor , Robert Tibshirani, and Jerome H. Friedman (2009). The Elements of Statistical
Learning: Data Mining, Inference, and Prediction (2 ed.). New York: Springer.
Herron, Michael C. (2019). Mail-in absentee ballot anomalies in north carolina’s 9th congressional
Herron, Michael C. and Daniel A. Smith (2012). Souls to the polls: Early voting in florida in the
Hubler, Shawn (2022, September 16). The California County Where MAGA Took Control.
Kang, Cecilia (2022). The Most Dominant Toxic Election Narratives Online. The New York
Times. https://www.nytimes.com/2022/09/23/technology/midterm-elections-misinformation.
Levitt, Justin (2007). The Truth About Voter Fraud. Technical report, Brennan Center for Justice
Li, Yimeng , Michelle Hyun, and R. Michael Alvarez (2022). Why do election results change after
election day? the “blue shift” in california elections. Political Research Quarterly 75 (3), 860–874.
45
Marks, Joseph (2021, August 11). The Cybersecurity 202: My Pillow cyber symposium is
yet another font of election fraud lies. The Washington Post. https://www.washingtonpost.
com/politics/2021/08/11/cybersecurity-202-my-pillow-cyber-symposium-is-yet-another-font-
Mazo, Eugene D. (2018). Finding Common Ground on Voter ID Laws. Univeristy of Memphis Law
Minnite, Lorraine (2007). Election Day Registration: A Study of Voter Fraud Allegations and
Minnite, Lorraine C. (2010). The Myth of Voter Fraud. Ithaca, NY: Cornell University.
Mollenkamp, Allison , Miles Parks, and Nick McMillan (2022). Election deniers
Murray, Sara and Jeff Simon (2022, January 29). The 2020 election wasn’t stolen. but douglas frank
and his bogus equation claiming otherwise are still winning over audiences. CNN.com. https://
2022.
National Conference of State Legislatures (2022). Access to and use of voter registration
lists. https://www.ncsl.org/research/elections-and-campaigns/access-to-and-use-of-voter-
Placer County Elections Office (2020). Placer county 2020 general election results. https://www.
Robison, Mark (2022, March 14). Robert Beadles is wealthy, new to town and wants to upend
46
03/14/robert-beadles-california-washoe-county-politics/9419737002/ (last accessed October 12,
2022).
Robison, Mark and Rio Lacanlale (2022, July 7). Lombardo retains win over
Gilbert as Washoe, Clark certify recount of GOP governor primary. Reno Gazette-
Journal. https://www.rgj.com/story/news/2022/06/16/joey-gilbert-pro-trump-ally-refuses-
12, 2022).
Rumble (2022, August 10). Gilbert vs Lombardo, Hearing for Summary Judgement. https://
Scherer, Zachary (2021, April 29). What Methods Did People Use to Vote in the 2020 Election?
Seeman, Matthew (2022, September 22). Judge orders sanctions against Joey Gilbert in failed
sanctions-against-joey-gilbert-in-failed-lawsuit-over-gop-primary-nevada-governor-election-
Seidman, Andrew (2021, July 26). Kathy Barnette’s futile hunt for voter fraud outside
Sides, John , Michael Tesler, and Lynn Vavreck (2019). Identity Crisis: The 2016 Presidential
Campaign and the Battle for the Meaning of America (2nd ed.). Princeton, NJ: Princeton
University Press.
47
Wu, Jennifer , Chenoa Yorgason, Hanna Folsz, Sandy Handan-Nader, Andrew Myers, Toby
Nowacki, Daniel M. Thompson, Jesse Yoder, and Andrew B. Hall (2020). Are Dead People Voting
By Mail? Evidence From Washington State Administrative Records. Unpublished working paper.
Yoder, Jesse , Cassandra Handan-Nader, Andrew Myers, Tobias Nowacki, Daniel M. Thompson,
Jennifer A. Wu, Chenoa Yorgason, and Andrew B. Hall (2021). How did absentee voting affect
48
A Solomon’s transformations mechanically induce
2
high R values
The “extraordinarily high” relationship between the variables g, h, and α — which the
Gilbert contest interprets as evidence of vote manipulation — is induced mechanically as a
result of Solomon’s transformations of the variables a - d, summarized in Equations 1 - 3. As
evidence of this, in this appendix we describe a set of simulation showing that, even when
underlying vote counts a - d are generated without any illicit manipulation, the regression in
2
Equation 4 returns a high value of R . This should not be considered surprising insofar as
the same variables (a - d) appear in both the regression’s left-hand side and its right-hand
side.
We offer three versions of our simulation, which have qualitatively identical results. Each
simulation is based on a set of 814 precincts, which mimics Clark County as of November
2022.
In our first simulation (“Poisson”), for each precinct we draw mail ballot turnout from a
(rounded) uniform distribution on the interval [100, 300] and we draw early voting turnout
similarly and independently. For each precinct, we then draw a from a Poisson distribution
with rate parameter equal to one-half of mail turnout and c from a Poisson distribution with
rate parameter equal to one-half of early voting turnout. We then set b equal to the difference
between mail ballot turnout and a and likewise set d equal to the difference between early
voting turnout and c. Finally, we transform our our variables a - d using Equations 1 - 3
2
and then estimate Equation 4, recording the value of R .
In our second simulation (“Uniform”), for each of our 814 precincts we draw mail ballot
turnout and early voting turnout as described in the first simulation immediately above.
We then draw a from a (rounded) uniform distribution bounded to the left by one and
to the right by mail ballot turnout. And, we similarly draw c from a (rounded) uniform
distribution bounded to the left by one and to the right by early voting turnout. As in the
first simulation, we then set b equal to the difference between mail ballot turnout and a
49
and likewise set d equal to the difference between early voting turnout and c. Finally, we
transform our randomly drawn a - d using Equations 1 - 3 and then estimate Equation 4,
2
recording the value of R .
In our third simulation (“beta”), for each of our 814 precincts we draw mail ballot
turnout and early voting turnout as described in the first simulation. We then specify a for
each precinct by multiplying mail ballot turnout with a randomly drawn value from a beta
distribution whose shape parameters are the maximum likelihood estimates produced by
fitting a beta density to Lombardo’s mail-in ballot support rate among precincts in Clark
County in the 2022 Republican gubernatorial primary. We similarly specify c for each
simulated precinct by multiplying early voting turnout with a randomly drawn value from
a beta distribution whose shape parameters are the maximum likelihood estimates
produced by fitting a beta density to Lombardo’s observed early voting support rate from
18
precincts in Clark County. As in the first simulation, we then set b equal to the difference
between mail ballot turnout and a and likewise set d equal to the difference between early
voting turnout and c. Finally, we transform our randomly drawn a - d using Equations 1 -
2
3 and then estimate Equation 4, recording the value of R .
2
Figure A.1 displays histograms of R values from our three simulations, each of which
was run 5,000 times. In all cases, these values are close to one. In the “Poisson” simulation,
2 2
the average R was 0.98, and the simulated R values were between 0.95 and one. In the
2
“Uniform” simulation, the average R value was 0.85 and in the “Beta” simulation, 0.94.
None of our simulations incorporated any form of vote manipulation. Nonetheless, what
we see from this histograms in Figure A.1 is that the transformation of Daugherity’s variables
2
a - d, specified in Equations 1 - 3, induces a high R value in a regression of g on h and α.
2
Daugherity interprets the high R value he observes as evidence of vote manipulation, but
2
our simulations make clear that in fact the R value noted by Daugherity is in fact evidence
18
When we fit beta distributions to Lombardo’s mail ballot and early voting support rates by Clark County
precinct, we disregard rates that are zero, one, or undefined on account of zero mail or early voting ballots
having been cast.
50
100 150 200 250 300
30
25
Density
Density
20
15
10
50
5
0
0
0.80 0.85 0.90 0.95 1.00 0.80 0.85 0.90 0.95 1.00
R2 R2
40
20
0
R2
of nothing.
51
B Frank’s “phantom voters” are nothing more than
voter file maintenance
A key feature of Frank’s claims of vote manipulation is that voter rolls are contaminated
with “phantom voters” who can be deployed to overturn election results. Frank asserts that
these voters serve as turnout credits that can be used to determine election outcomes.
The basis for Frank’s claims about phantom voters rests upon discrepancies between
statewide voter files — which we previously noted are lists of registered voters in a state —
and official election results. Drawing on voter files, proponents of the idea that phantom
voters exist point out that the total number of voters in any particular voter file who are
listed as having turned out to vote in a given election tends to be less than official voter
turnout in said election. One obvious explanation for this is that voters are removed from
voter rolls because they have moved, died, or otherwise canceled their registrations. When a
voter is removed from a voter file, this voter will no longer be counted when using the voter
file to calculate overall turnout in elections in which the voter participated. Over time, for
any given election in a state this will lead to the state’s voter file having fewer voters who
participated in said election than actually voted in it.
For example, in public testimony in Wisconsin, Frank Frank argued on December 8,
2021 that drops in the number of registered voters are a surprising feature of the Wisconsin
19
voter registration system. In fact, Frank appears unaware that such drops coincide with
efforts to identify voters who have moved. Rather than reasonably attributing differences
between counts of voter file voters and official turnout statistics to voter file maintenance,
Frank argues that these these differences reflect fictitious voters added to the rolls so as to
determine election outcomes.
To assess his phantom voter theory, Frank runs non-random surveys, which he calls
“canvassing,” to identify what he deems to be suspicious activity (e.g., Frank 2021a). In
19
The Wisconsin legislature has records of Frank’s testimony at https://legis.wisconsin.gov/assembly/59/
ramthun/media/1348/dr-frank-12-8-2021-wisconsin.pdf (last accessed October 24, 2022).
52
particular, it is our understanding that Frank coordinates with local groups to go
door-to-door in neighborhoods to assess discrepancies between official vote records and the
recollections of residents. While, to the best of our knowledge Frank has not presented
official statistics on individuals he believes did not vote in a given election but are recorded
as having voted in it, he has claimed in presentations that he uncovers individuals who fit
this description as well as residents who claim that individuals recorded as voting in the
election do not live at their supposed addresses. Frank has not addressed alternative
explanations for the patterns he thinks are evidence of malfeasance — including the
possibility that voters may fail to recall the elections in which they have, that residents of
multi-unit buildings may not know all residents of the building, and that poorly trained
canvassers may make errors when identifying the current residents of a building.
To better understand who is removed from the voter rolls and the extent to which they
are actually real individuals, we take advantage of a policy established in Georgia that calls
for maintaining lists of all individuals who vote by mail, i.e., absentee voter lists. These
lists are not updated when voters are removed from the voter rolls. Using unique Georgia
voter identification numbers, we can link individuals from, say, a current Georgia voter file
to absentee voting list from prior elections. After identifying individuals who were removed
from a voter file (these individuals are on an absentee voter list but not in said voter file),
we can engage in a simple search to confirm that they are real individuals.
In the 2020 presidential election in Georgia, the total number of voters who cast a vote
20
for president was 4,998,482, and total turnout in the election was 5,023,159. The number of
voters in the December 2020 voter file, according to L2 was 4,905,383 a difference of 117,776
(about 2.3 percent of the total votes cast). We then used the absentee voter file to determine
if we could identify who was removed from the file. Of those who are removed, we were able
20
For total presidential votes cast in the 2020 general election in Georgia, see data produced by the
Georgia Secretary of State, available at https://results.enr.clarityelections.com/GA/105369/web.264614/#/
summary (last accessed October 24, 2022). For total turnout, we downloaded official Georgia turnout by
demographic data at https://sos.ga.gov/sites/default/files/bulk/Voter Turn Out By Demographics.zip (last
accessed October 24, 2022), extracted the file for the 2020 general election, and summed column BN.
53
to find 66,863 who had cast absentee ballots or approximately 56.8 percent. Simple internet
searches reveal it is clear that these are real individuals. In fact, we can obtain records of
individuals who moved after the election and therefore changed their voter registrations. In
short, there is little reason to suspect that individuals who voted in the 2020 election but do
not appear in the December 2020 Georgia voter file are phantom voters.
The absentee ballot records maintained by the Georgia Secretary of State also show
the Georgia system canceling the ballots of individuals who are no longer eligible to vote.
For example, in our searches we were able to identify an individual who had requested an
absentee ballot for the 2020 general election before passing away before it. This individual’s
21
ballot was canceled through an administrative process.
In short, a simple analysis of a state, Georgia, that provides information on who is
removed from its voter reveals no evidence of nefarious “phantom voters.” Instead, what it
shows is a state where election officials update registration records and seek to block deceased
individuals from receiving ballots.
21
To avoid revealing personal information, we are not publicizing names of such individuals. We are able
to discuss these names with interested researchers.
54
C Emails to Douglas Frank for replication code and
providing memo
55
D Additional turnout figures for Placer County
This appendix contains two histograms describing Democratic and overall voter turnout
rates, by precinct, for Placer County in the 2020 general election. As both plots in the figure
make clear, there was considerable variability in these turnout rates in November 2020.
Figure D.2: Democratic and overall turnout rates in Placer County in the 2020 general
election.
56
E L2 statewide voter files
In the body of the paper, we described how we used statewide voter files from L2 to
characterize age-based voter turnout by county and state. Table 4 in this appendix lists the
dates on which the 42 voter files were pulled.
57